2 * Copyright (c) 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_shared_reference.cpp
27 * IR lower pass to replace dereferences of compute shader shared variables
28 * with intrinsic function calls.
30 * This relieves drivers of the responsibility of allocating space for the
31 * shared variables in the shared memory region.
34 #include "lower_buffer_access.h"
35 #include "ir_builder.h"
36 #include "main/macros.h"
37 #include "util/list.h"
38 #include "glsl_parser_extras.h"
40 using namespace ir_builder
;
45 struct list_head node
;
46 const ir_variable
*var
;
50 class lower_shared_reference_visitor
:
51 public lower_buffer_access::lower_buffer_access
{
54 lower_shared_reference_visitor(struct gl_linked_shader
*shader
)
55 : list_ctx(ralloc_context(NULL
)), shader(shader
), shared_size(0u)
57 list_inithead(&var_offsets
);
60 ~lower_shared_reference_visitor()
62 ralloc_free(list_ctx
);
71 void insert_buffer_access(void *mem_ctx
, ir_dereference
*deref
,
72 const glsl_type
*type
, ir_rvalue
*offset
,
73 unsigned mask
, int channel
);
75 void handle_rvalue(ir_rvalue
**rvalue
);
76 ir_visitor_status
visit_enter(ir_assignment
*ir
);
77 void handle_assignment(ir_assignment
*ir
);
79 ir_call
*lower_shared_atomic_intrinsic(ir_call
*ir
);
80 ir_call
*check_for_shared_atomic_intrinsic(ir_call
*ir
);
81 ir_visitor_status
visit_enter(ir_call
*ir
);
83 unsigned get_shared_offset(const ir_variable
*);
85 ir_call
*shared_load(void *mem_ctx
, const struct glsl_type
*type
,
87 ir_call
*shared_store(void *mem_ctx
, ir_rvalue
*deref
, ir_rvalue
*offset
,
91 struct gl_linked_shader
*shader
;
92 struct list_head var_offsets
;
98 lower_shared_reference_visitor::get_shared_offset(const ir_variable
*var
)
100 list_for_each_entry(var_offset
, var_entry
, &var_offsets
, node
) {
101 if (var_entry
->var
== var
)
102 return var_entry
->offset
;
105 struct var_offset
*new_entry
= rzalloc(list_ctx
, struct var_offset
);
106 list_add(&new_entry
->node
, &var_offsets
);
107 new_entry
->var
= var
;
109 unsigned var_align
= var
->type
->std430_base_alignment(false);
110 new_entry
->offset
= glsl_align(shared_size
, var_align
);
112 unsigned var_size
= var
->type
->std430_size(false);
113 shared_size
= new_entry
->offset
+ var_size
;
115 return new_entry
->offset
;
119 lower_shared_reference_visitor::handle_rvalue(ir_rvalue
**rvalue
)
124 ir_dereference
*deref
= (*rvalue
)->as_dereference();
128 ir_variable
*var
= deref
->variable_referenced();
129 if (!var
|| var
->data
.mode
!= ir_var_shader_shared
)
132 buffer_access_type
= shared_load_access
;
134 void *mem_ctx
= ralloc_parent(shader
->ir
);
136 ir_rvalue
*offset
= NULL
;
137 unsigned const_offset
= get_shared_offset(var
);
140 assert(var
->get_interface_type() == NULL
);
141 const enum glsl_interface_packing packing
= GLSL_INTERFACE_PACKING_STD430
;
143 setup_buffer_access(mem_ctx
, deref
,
144 &offset
, &const_offset
,
145 &row_major
, &matrix_columns
, NULL
, packing
);
147 /* Now that we've calculated the offset to the start of the
148 * dereference, walk over the type and emit loads into a temporary.
150 const glsl_type
*type
= (*rvalue
)->type
;
151 ir_variable
*load_var
= new(mem_ctx
) ir_variable(type
,
154 base_ir
->insert_before(load_var
);
156 ir_variable
*load_offset
= new(mem_ctx
) ir_variable(glsl_type::uint_type
,
157 "shared_load_temp_offset",
159 base_ir
->insert_before(load_offset
);
160 base_ir
->insert_before(assign(load_offset
, offset
));
162 deref
= new(mem_ctx
) ir_dereference_variable(load_var
);
164 emit_access(mem_ctx
, false, deref
, load_offset
, const_offset
, row_major
,
165 matrix_columns
, packing
, 0);
173 lower_shared_reference_visitor::handle_assignment(ir_assignment
*ir
)
178 ir_rvalue
*rvalue
= ir
->lhs
->as_rvalue();
182 ir_dereference
*deref
= ir
->lhs
->as_dereference();
186 ir_variable
*var
= ir
->lhs
->variable_referenced();
187 if (!var
|| var
->data
.mode
!= ir_var_shader_shared
)
190 buffer_access_type
= shared_store_access
;
192 /* We have a write to a shared variable, so declare a temporary and rewrite
193 * the assignment so that the temporary is the LHS.
195 void *mem_ctx
= ralloc_parent(shader
->ir
);
197 const glsl_type
*type
= rvalue
->type
;
198 ir_variable
*store_var
= new(mem_ctx
) ir_variable(type
,
201 base_ir
->insert_before(store_var
);
202 ir
->lhs
= new(mem_ctx
) ir_dereference_variable(store_var
);
204 ir_rvalue
*offset
= NULL
;
205 unsigned const_offset
= get_shared_offset(var
);
208 assert(var
->get_interface_type() == NULL
);
209 const enum glsl_interface_packing packing
= GLSL_INTERFACE_PACKING_STD430
;
211 setup_buffer_access(mem_ctx
, deref
,
212 &offset
, &const_offset
,
213 &row_major
, &matrix_columns
, NULL
, packing
);
215 deref
= new(mem_ctx
) ir_dereference_variable(store_var
);
217 ir_variable
*store_offset
= new(mem_ctx
) ir_variable(glsl_type::uint_type
,
218 "shared_store_temp_offset",
220 base_ir
->insert_before(store_offset
);
221 base_ir
->insert_before(assign(store_offset
, offset
));
223 /* Now we have to write the value assigned to the temporary back to memory */
224 emit_access(mem_ctx
, true, deref
, store_offset
, const_offset
, row_major
,
225 matrix_columns
, packing
, ir
->write_mask
);
231 lower_shared_reference_visitor::visit_enter(ir_assignment
*ir
)
233 handle_assignment(ir
);
234 return rvalue_visit(ir
);
238 lower_shared_reference_visitor::insert_buffer_access(void *mem_ctx
,
239 ir_dereference
*deref
,
240 const glsl_type
*type
,
245 if (buffer_access_type
== shared_store_access
) {
246 ir_call
*store
= shared_store(mem_ctx
, deref
, offset
, mask
);
247 base_ir
->insert_after(store
);
249 ir_call
*load
= shared_load(mem_ctx
, type
, offset
);
250 base_ir
->insert_before(load
);
251 ir_rvalue
*value
= load
->return_deref
->as_rvalue()->clone(mem_ctx
, NULL
);
252 base_ir
->insert_before(assign(deref
->clone(mem_ctx
, NULL
),
258 compute_shader_enabled(const _mesa_glsl_parse_state
*state
)
260 return state
->stage
== MESA_SHADER_COMPUTE
;
264 lower_shared_reference_visitor::shared_store(void *mem_ctx
,
269 exec_list sig_params
;
271 ir_variable
*offset_ref
= new(mem_ctx
)
272 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
273 sig_params
.push_tail(offset_ref
);
275 ir_variable
*val_ref
= new(mem_ctx
)
276 ir_variable(deref
->type
, "value" , ir_var_function_in
);
277 sig_params
.push_tail(val_ref
);
279 ir_variable
*writemask_ref
= new(mem_ctx
)
280 ir_variable(glsl_type::uint_type
, "write_mask" , ir_var_function_in
);
281 sig_params
.push_tail(writemask_ref
);
283 ir_function_signature
*sig
= new(mem_ctx
)
284 ir_function_signature(glsl_type::void_type
, compute_shader_enabled
);
286 sig
->replace_parameters(&sig_params
);
287 sig
->is_intrinsic
= true;
288 sig
->intrinsic_id
= ir_intrinsic_shared_store
;
290 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_store_shared");
291 f
->add_signature(sig
);
293 exec_list call_params
;
294 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
295 call_params
.push_tail(deref
->clone(mem_ctx
, NULL
));
296 call_params
.push_tail(new(mem_ctx
) ir_constant(write_mask
));
297 return new(mem_ctx
) ir_call(sig
, NULL
, &call_params
);
301 lower_shared_reference_visitor::shared_load(void *mem_ctx
,
302 const struct glsl_type
*type
,
305 exec_list sig_params
;
307 ir_variable
*offset_ref
= new(mem_ctx
)
308 ir_variable(glsl_type::uint_type
, "offset_ref" , ir_var_function_in
);
309 sig_params
.push_tail(offset_ref
);
311 ir_function_signature
*sig
=
312 new(mem_ctx
) ir_function_signature(type
, compute_shader_enabled
);
314 sig
->replace_parameters(&sig_params
);
315 sig
->is_intrinsic
= true;
316 sig
->intrinsic_id
= ir_intrinsic_shared_load
;
318 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_load_shared");
319 f
->add_signature(sig
);
321 ir_variable
*result
= new(mem_ctx
)
322 ir_variable(type
, "shared_load_result", ir_var_temporary
);
323 base_ir
->insert_before(result
);
324 ir_dereference_variable
*deref_result
= new(mem_ctx
)
325 ir_dereference_variable(result
);
327 exec_list call_params
;
328 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
330 return new(mem_ctx
) ir_call(sig
, deref_result
, &call_params
);
333 /* Lowers the intrinsic call to a new internal intrinsic that swaps the access
334 * to the shared variable in the first parameter by an offset. This involves
335 * creating the new internal intrinsic (i.e. the new function signature).
338 lower_shared_reference_visitor::lower_shared_atomic_intrinsic(ir_call
*ir
)
340 /* Shared atomics usually have 2 parameters, the shared variable and an
341 * integer argument. The exception is CompSwap, that has an additional
344 int param_count
= ir
->actual_parameters
.length();
345 assert(param_count
== 2 || param_count
== 3);
347 /* First argument must be a scalar integer shared variable */
348 exec_node
*param
= ir
->actual_parameters
.get_head();
349 ir_instruction
*inst
= (ir_instruction
*) param
;
350 assert(inst
->ir_type
== ir_type_dereference_variable
||
351 inst
->ir_type
== ir_type_dereference_array
||
352 inst
->ir_type
== ir_type_dereference_record
||
353 inst
->ir_type
== ir_type_swizzle
);
355 ir_rvalue
*deref
= (ir_rvalue
*) inst
;
356 assert(deref
->type
->is_scalar() && deref
->type
->is_integer());
358 ir_variable
*var
= deref
->variable_referenced();
361 /* Compute the offset to the start if the dereference
363 void *mem_ctx
= ralloc_parent(shader
->ir
);
365 ir_rvalue
*offset
= NULL
;
366 unsigned const_offset
= get_shared_offset(var
);
369 assert(var
->get_interface_type() == NULL
);
370 const enum glsl_interface_packing packing
= GLSL_INTERFACE_PACKING_STD430
;
371 buffer_access_type
= shared_atomic_access
;
373 setup_buffer_access(mem_ctx
, deref
,
374 &offset
, &const_offset
,
375 &row_major
, &matrix_columns
, NULL
, packing
);
379 assert(matrix_columns
== 1);
381 ir_rvalue
*deref_offset
=
382 add(offset
, new(mem_ctx
) ir_constant(const_offset
));
384 /* Create the new internal function signature that will take an offset
385 * instead of a shared variable
387 exec_list sig_params
;
388 ir_variable
*sig_param
= new(mem_ctx
)
389 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
390 sig_params
.push_tail(sig_param
);
392 const glsl_type
*type
= deref
->type
->base_type
== GLSL_TYPE_INT
?
393 glsl_type::int_type
: glsl_type::uint_type
;
394 sig_param
= new(mem_ctx
)
395 ir_variable(type
, "data1", ir_var_function_in
);
396 sig_params
.push_tail(sig_param
);
398 if (param_count
== 3) {
399 sig_param
= new(mem_ctx
)
400 ir_variable(type
, "data2", ir_var_function_in
);
401 sig_params
.push_tail(sig_param
);
404 ir_function_signature
*sig
=
405 new(mem_ctx
) ir_function_signature(deref
->type
,
406 compute_shader_enabled
);
408 sig
->replace_parameters(&sig_params
);
409 sig
->is_intrinsic
= true;
411 assert(ir
->callee
->intrinsic_id
>= ir_intrinsic_generic_load
);
412 assert(ir
->callee
->intrinsic_id
<= ir_intrinsic_generic_atomic_comp_swap
);
413 sig
->intrinsic_id
= MAP_INTRINSIC_TO_TYPE(ir
->callee
->intrinsic_id
, shared
);
416 sprintf(func_name
, "%s_shared", ir
->callee_name());
417 ir_function
*f
= new(mem_ctx
) ir_function(func_name
);
418 f
->add_signature(sig
);
420 /* Now, create the call to the internal intrinsic */
421 exec_list call_params
;
422 call_params
.push_tail(deref_offset
);
423 param
= ir
->actual_parameters
.get_head()->get_next();
424 ir_rvalue
*param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
425 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
426 if (param_count
== 3) {
427 param
= param
->get_next();
428 param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
429 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
431 ir_dereference_variable
*return_deref
=
432 ir
->return_deref
->clone(mem_ctx
, NULL
);
433 return new(mem_ctx
) ir_call(sig
, return_deref
, &call_params
);
437 lower_shared_reference_visitor::check_for_shared_atomic_intrinsic(ir_call
*ir
)
439 exec_list
& params
= ir
->actual_parameters
;
441 if (params
.length() < 2 || params
.length() > 3)
445 ((ir_instruction
*) params
.get_head())->as_rvalue();
449 ir_variable
*var
= rvalue
->variable_referenced();
450 if (!var
|| var
->data
.mode
!= ir_var_shader_shared
)
453 const char *callee
= ir
->callee_name();
454 if (!strcmp("__intrinsic_atomic_add", callee
) ||
455 !strcmp("__intrinsic_atomic_min", callee
) ||
456 !strcmp("__intrinsic_atomic_max", callee
) ||
457 !strcmp("__intrinsic_atomic_and", callee
) ||
458 !strcmp("__intrinsic_atomic_or", callee
) ||
459 !strcmp("__intrinsic_atomic_xor", callee
) ||
460 !strcmp("__intrinsic_atomic_exchange", callee
) ||
461 !strcmp("__intrinsic_atomic_comp_swap", callee
)) {
462 return lower_shared_atomic_intrinsic(ir
);
469 lower_shared_reference_visitor::visit_enter(ir_call
*ir
)
471 ir_call
*new_ir
= check_for_shared_atomic_intrinsic(ir
);
474 base_ir
->replace_with(new_ir
);
475 return visit_continue_with_parent
;
478 return rvalue_visit(ir
);
481 } /* unnamed namespace */
484 lower_shared_reference(struct gl_linked_shader
*shader
, unsigned *shared_size
)
486 if (shader
->Stage
!= MESA_SHADER_COMPUTE
)
489 lower_shared_reference_visitor
v(shader
);
491 /* Loop over the instructions lowering references, because we take a deref
492 * of an shared variable array using a shared variable dereference as the
493 * index will produce a collection of instructions all of which have cloned
494 * shared variable dereferences for that array index.
498 visit_list_elements(&v
, shader
->ir
);
499 } while (v
.progress
);
501 *shared_size
= v
.shared_size
;